Fix Cohere compatibility logit scaling

[emerardd]

Description

Fixes #1726

Cohere and Cohere2 weight processing folds cfg.logit_scale into the live unembedding weights, but the Hugging Face model forward still applied its model-level scale afterward. As a result, the first valid compatibility-mode conversion produced logits with the scale applied twice.

This change adds an architecture post-install lifecycle hook. The Cohere adapter uses it to neutralize the Hugging Face model-level logit_scale only after the scaled unembedding weights have been installed. The Bridge config retains the original scale as metadata, and tied input embeddings remain unscaled.

Network-free tiny-config regressions cover both direct process_weights() and the first enable_compatibility_mode() call.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Validation

• tests/integration/model_bridge/test_cohere_adapter.py: 24 passed
• Cohere/Cohere2 adapter unit surface: 74 passed
• Weight-processing unit surface: 27 passed, 1 skipped
• Full non-slow unit surface: 5,435 passed; the sole nonzero result was a native-Windows GBK decode error in a source-roster test
• The exact roster test passed after enabling Python UTF-8 mode: 1 passed
• mypy .: Success, no issues found in 392 source files
• pycln, pinned isort, Black, and git diff --check: passed

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation (no public documentation change is required; the internal lifecycle hook is documented in code)
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Thanks for documenting this issue and putting together the fix @emerardd! Just a note on the CI failure, that is a testing issue that is currently being resolved in #1730, and should be ready soon. A couple small comments on this fix, let me know if you have any questions

[jlarson4]

supported_architectures/AGENTS.md "When to override preprocess_weights" tells adapter authors to fold any config knob HF applies in forward, with Cohere's logit_scale as its worked example and a copy-paste skeleton. Following it produces exactly this bug, because it never says the forward-time factor also has to be taken out.

Please add postprocess_weights to that section: when the factor lives outside the module the bridge wraps, folding it is only half the change.

[emerardd]

Addressed in 1544c4ce. I expanded the preprocess_weights guidance in supported_architectures/AGENTS.md to cover the paired post-install lifecycle: when HF applies a factor outside the wrapped component, adapters now track whether the fold occurred and use postprocess_weights() to neutralize the outer runtime factor after installation. The guidance and skeleton also call out that a missing runtime attribute means there is no outer factor to neutralize.

[jlarson4]

If original_model has no logit_scale, its forward isn't applying one, so the fold alone is already correct and there is nothing to neutralize. A Cohere-architecture module in that shape processes cleanly on dev and hard-fails here. This occurs after the folded weights are installed and _weights_processed is set, so a caller that catches the error is left holding the mis-scaled bridge.

Return instead of raising when the attribute is absent.

[emerardd]

Addressed in 1544c4ce. postprocess_weights() now neutralizes model.logit_scale only when the wrapped model exposes it, and always clears the pending fold state without raising when the attribute is absent. I added a network-free regression that removes the model-level attribute, runs process_weights(), and verifies that processing completes, the attribute remains absent, the unembed weight is folded by 0.5, and cfg.logit_scale remains 0.5. Focused validation passed: 3 end-to-end regressions, 74 Cohere/Cohere2 adapter tests, 27 weight-processing tests with 1 skip, formatting, and mypy over 392 source files.